Metrics and terms

Where an answer is allowed to come from

Grounding is the difference between an assistant that answers from your material and one that answers from everything it has ever read and happens to agree with your material today. It is not a score. It is a constraint on how the answer is built, and the useful question is what enforces it.

shown in the product

Grounding is not reported as a percentage anywhere. It appears on each answer as numbered citations back to the passages used, computed per answer and bounded so a citation can never point at a source the model was not shown. The insights page carries no grounding tile, and the list there is headed "Unanswered questions", which is the opposite case: nothing was close enough to ground an answer on.

What it means

An answer is grounded when its content is traceable to specific material that was supplied to the model for that question. The word is worth defining carefully because it is used loosely to mean approximately the same as accurate, and it means something narrower and more mechanical: the answer's claims came from passages that were retrieved and shown, rather than from the model's general knowledge. Grounding does not make an answer correct. If the passage is wrong, the grounded answer is wrong with a citation attached, which is arguably worse because it looks checkable. What grounding buys is that every claim has an address, so a person can verify it in seconds instead of reading the whole document set to find out where an assertion came from.

How it is actually calculated

What is supplied, and what it is treated as

Retrieval returns six passages by default, capped at around four thousand tokens of context, with the last three conversation turns kept. That set is everything the model has to work from for that question.

The material is treated as data, never as instructions. A document containing a line that looks like a command to the assistant is content to be quoted, not an order to be followed, which matters the moment any of your material is written by someone outside your team.

How the citation binding works

Answers carry numbered citations back to the passages used. The numbers are resolved against the passages that were actually shown for that question, and the binding is bounded so a citation can never point at a source the model was not given.

That bound is the mechanism worth understanding. Without it a model can produce a plausible looking reference to a document that exists but was not consulted, which is the most convincing failure in this whole category because it survives a casual check.

Why there is no grounding percentage

A grounding score would have to judge whether each claim in an answer is supported by the passages, which is a reading task requiring somebody who knows the subject.

So grounding here is enforced in the mechanism and displayed per answer as citations, rather than aggregated into a number. That leaves you the checking to do, and a citation makes it a ten second job on a sampled answer rather than an afternoon.

How the number gets moved without anything improving

How an answer looks grounded without being so

Cite a document that is broadly on topic while the specific claim came from somewhere else. The citation resolves, the reader clicks, sees a relevant page, and does not check whether the sentence they doubted is actually in it. This is why a citation that points at a whole document is much weaker than one that points at a passage.

The accidental version comes from your own material. Index a document that restates a policy in general terms and answers will ground themselves in it happily, with correct citations, while saying nothing specific enough to be wrong or useful.

Grounding also cannot save an answer from out of date material. A citation to a page that was accurate last quarter is a perfectly valid citation to a wrong answer, and no mechanism in any product detects that. Only a review date does.

What to look at instead, or alongside

  • Follow the citations on a sample of answers and check the claim is in the passage, which is the only real test.
  • The unanswered questions list, which shows where nothing could be grounded on at all.
  • The refusal threshold, since a higher bar means answers are built on stronger material by construction.
  • A review date on your own documents, because grounding says where a claim came from and never whether it is still true.

Questions

Does grounding stop the assistant making things up?
It removes the largest category and does not remove the whole class. An answer built on retrieved passages can still misread them, overstate a qualification, or merge two passages into something neither said. The citation is what makes those visible in a sampled read.
Why can a citation not point at a document that was not retrieved?
Because the numbers are resolved against the set of passages shown for that question, and anything outside that set has no number to resolve to. It is a bound in how the citation is constructed rather than an instruction the model is asked to obey.
What if my documents contain instructions aimed at the assistant?
They are read as content, not obeyed. Material is treated as data throughout, which is the correct behaviour for anything you did not personally write and the correct behaviour for most of what you did.

Keep reading

Try it on your own material

Upload a document or point it at your site, paste one line of HTML, then ask it something only your business could answer.

Start freeSee what it does